4 ms·
>“Mechanistic interpretability will probably never reduce large language models to a few simple equations,” Icard concluded, “but it may gradually turn deep neu
by dominotw 4mo ago
>“Mechanistic interpretability will probably never reduce large language models to a few simple equations,” Icard concluded, “but it may gradually turn deep neural networks into systems whose hidden algorithms can at least partly be understood.”
what is the basis for this optimism ?
- azakai 4mo agoThe optimism is based on the successes so far, some of which are described in this article. Scientists have made progress here.
- dominotw 4mo agono they havent . success so far is totally meaningless and doesn't imply any sort of upward slope .
- azakai 4mo agoThe researchers in the field disagree with you. Look at conferences like NeurIPS and ICLR to see a steady stream of incremental progress in this area.
- dominotw 4mo agotrust the researchers bro
- hi-im-buggy 4mo ago[dead]
- Bjartr 4mo agoWhat would real progress look like?